Shared Task on Multimodal Hate and Sentiment Understanding in Low-Resource Memes@CHiPSAL2026
Memes in low-resource languages such as Nepali play an increasingly important role in online discourse, yet they remain significantly underrepresented in multimodal research. This shared task focuses on multimodal hate detection in Nepali memes, where each meme is annotated as hate or no-hate. The task challenges participants to develop models that can effectively integrate visual and textual cues to detect harmful content in a low-resource linguistic setting marked by cultural context, code-mixing, sarcasm, and informal language. By addressing the scarcity of annotated resources and benchmarks for Nepali, this challenge aims to push forward fair, inclusive, and robust multimodal hate detection systems that generalize beyond high-resource languages.
Subtask A: Hate Speech Detection
This subtask is a binary classification problem where systems must determine whether a Nepali meme contains hate speech or not. Participants must build multimodal models capable of labeling each meme as hate or non-hate.
This task focuses on identifying harmful or abusive content in Nepali memes. The challenge lies in the interplay between visual symbolism, culturally embedded humor, informal expressions, and textual cues. Since Nepali is a low-resource language with limited annotated datasets, this task encourages approaches that integrate cross-modal learning, cultural context modeling, and transfer learning. Systems will be evaluated on their ability to capture explicit and implicit aggression, sarcasm, and targeting that appears in online discourse.
Subtask A in codabench: https://www.codabench.org/competitions/12090/
This subtask is a three-way classification problem where systems must determine whether the sentiment of a Nepali meme is positive, negative, or neutral. Participants must develop multimodal models capable of assigning one of these sentiment labels to each meme.
This task addresses emotion and opinion understanding in Nepali meme content. Unlike standard text-only sentiment classification, memes involve visual metaphors, humor, irony, and non-literal cues that complicate interpretation. Combined with code mixing and informal Nepali vernacular, effective sentiment analysis requires reasoning across both image and text. This subtask advances multimodal affect understanding in low-resource settings and evaluates how well systems capture nuanced sentiment beyond surface-level polarity.
Subtask B in Codabench: https://www.codabench.org/competitions/12091/
IMPORTANT DATES
Start of the Competition: Dec 9, 2025
Eval Phase Start: Dec 9, 2025
Test Phase Start: Dec 25, 2025
Test Phase End: Feb 14, 2026
Paper Submission Deadline: Feb 20, 2026
Notification of acceptance: March 20, 2026
Camera Ready due: March 30, 2026
ORGANIZING TEAM
Surendrabikram Thapa (Virginia Tech, USA)
Shuvam Shiwakoti (Virginia Tech, USA)
Siddhant Bikram Shah (Northeastern University, USA)
Kritesh Rauniyar (Delhi Technological University, India)
Surabhi Adhikari (Columbia University, USA)
Kengatharaiyer Sarveswaran (University of Jaffna, Sri Lanka)
Kristina T. Johnson (Northeastern University, USA)
Bal Krishna Bal (Kathmandu University, Nepal)
Usman Naseem (Macquarie University, Australia)